Dreaming, in humans, is essentially a nightly re-training so that the most essential memories are encoded in weights as opposed to runtime additions.
Very curious to see how this works with Truffle, given the memory profile of the GPU under the hood.
2
27
The most dangerous outcome for AI is that it behaves like an ideologue.
Anthropic turns down people who are deeply talented technically.
Dario Amodei explained why to Jamie Dimon, CEO of JPMorgan, days before a $1.4 trillion IPO.
- Everyone goes through a culture interview
- No jerks, no one who tries to steal the show, no one indifferent to the public benefit
- He admits it's painful in the short run
- All 7 co-founders are still at the company
- More than 3 in 4 of the first 100 employees are still there
- Two senior product people quit and came back within two weeks
He spends 40-50% of his own time on talent and culture.
It was the first thing he chose to talk about in front of Wall Street.
Full breakdown here: theaiopportunities.com/p/ant…
14
My guess is that the extra time allows them to hire less (throughput management).
The check-in time at hotels is getting later and later (now it’s 2 pm in many places) and check-out is getting real early (10 am in some hotels).
What happened to 12-to-12?
You can’t take 4 hours to prepare a room. Seems to me a rip-off scheme.
15
This too, is a naive take if one considers all the actions by the government during the GFC.
Tyler Cowen thinks that if Anthropic leadership believed transformative AI was imminent, they wouldn’t IPO and give away part of the super valuable company they own.
I think that’s a naive take. While Anthropic is a private company, the trillion-plus dollars that is Anthropic is extremely vulnerable because it mostly belongs to a relatively small number of unlikable nerds in the Bay Area. If the federal government woke up and started taking the past statements of Anthropic leadership seriously, they could just immediately bring the hammer down and do very hardcore Situational Awareness type stuff, and the economic damage would be pretty contained.
Once Anthropic IPOs, damaging Anthropic means damaging a much broader-based coalition of ordinary American rich people, and suddenly there are a bunch of powerful stakeholders with a vested interest in preventing Anthropic from getting nationalized or otherwise wrecked.
I think that the IPO is exactly what you would expect from a company that plans to roll the dice soon and wants to make it as difficult as possible for anyone to stop them.
38
The Labs are furiously building the app layer (distribution) and the models are the 2020s version of $2 rides.
Models have no moat whatsoever. Zero.
Or pricing power.
Even the labs obviously know this, as they keep shutting models down and forcing people to switch to newer ones.
Every model is more like a new movie release than recurring profits. It’s basically a box office business.
1
41
Tangentially, if long-horizon simulations become reliable and affordable at scale, then AI might actually become superior to humans at strategy because it can simulate counterfactuals, something reality cannot provide ground truth on.
GTA 6 wasn’t meant for you to sit there twiddling your thumbs on cuckold simulator
It is a rich, and deep simulation environment for 27’ era agents
You will revisit this post next year as the takeoff in long horizon behavior stems from sim-like games
25
I would posit that it is the latter and is going to get worse.
India is going from having to import energy to importing energy and tokens.
Anthropic, an AI company known for Claude, listed at a valuation of $ 1.5 trillion. It did not exist till ~ 5 years ago. This is more than the combined market cap of the top 10 Indian companies. Are we undervalued, or do we need to catch up in the new world? Or both?
30
This slide is really accurate but the value from the gap won’t accrue to the labs. The real world is too messy.
A slide from @AnthropicAI's IPO roadshow. Hits hard and is very true.
AI is soooo much better than how most people are using it.
1
30
EXACTLY THIS.
AI will progress fastest where models training has a high number of shots on goal and where the outcome can be reliably measured.
AI research has this. Math has this. Coding has this. So it feels to the Labs that AGI is here.
Most other domains do not have this. AI diffusion is going to require a lot of engineering-savvy people rearchitecting the world.
A Ferrari is 10X faster than you on a racetrack, but you're infinitely faster than it crossing a jungle. The problem with the fast takeoff crowd in AI is that they think the world is a racetrack, when in fact it's more like a jungle.
1
22
Pre-training is *enormously* expensive (billions of dollars) and the version an Indian company would build first would likely be based on common crawl, meaning it would be a commodity.
So you’d want to acquire a pre-trained model. The Chinese models are based on and/or distilled from American pre-trained models, which are not available otherwise (smaller models are, but larger ones are not).
Basing the first few Indian models on these open weights models is a practical and cost-effective choice as long as there has been significant effort in tuning them to Indian tastes, needs, norms, and laws.
Sarvam got picked by the government in April 2025 to build India's own foundation model. They got a ₹98.68 crore GPU subsidy for 4,096 H100s and raised $234M at a $1.5B valuation in June.
Now they've opened Token Factory to everyone. 8 models on it. 1 is theirs. 6 are Chinese open weights (DeepSeek, Kimi, GLM) and 1 is Google's Gemma.
So India's sovereign AI lab is mostly serving Chinese models from Indian GPUs. That's a hosting business, and plenty of companies already run one.
I wanted Sarvam to be the lab that ships frontier research out of India. After that much capital and subsidized compute, this is not it.
51
Apple is tantalizingly close to having all the ingredients to dominate consumer AI, but getting there will take time.
* Edge hardware — they’ve been seeding the market for years, and newer iPhones can run ~3B models, but you probably need another order of magnitude capability to support truly edge inference
* iCloud — if not at the edge, a Muse-like architecture with user-level private inference might work, but it is still cost-prohibitive and not business-model compliant
* Harnesses (OS, app stores) — the core infrastructure for context is in place through the platform, this is the least risky dimension
* Product — Apple has built ML into its products very thoughtfully. They clearly understand how AI must be built for large scale consumer products
I think we are about 3-5 years away from edge capability and cloud-inference economics converging in Apple’s favor.
Local AI economics don’t work.
HOWEVER, a practical solution is to leverage local, less powerful models for agentic tasks and work that involves personal data; and outsource deep research tasks to the Labs.
The harness needs to be robust in terms of data tagging and extraction.
So, Local AI's economics is terrible, but by how much? I did some back-of-the-envelope calculations: assuming DeepSeek 4.1 Flash 1017, where input tokens is $0.15 per million, and output tokens (decode) is $0.6 per million. On M5 Ultra, you get 4000 tok/s prefill and 80 tok/s decode (conservative). Assuming 9:1 ratio, you are looking at making $45.36 per day. You are making the money back in about 9 months (~$12k for 256GiB variant). Again, this is not economical, because on NVIDIA hardware, you are making the money back in less than 4 months.
1
75

















